Dear Peter, On Sun, May 03, 2015 at 07:50:44AM +0200, Peter Bienstman wrote: > > It is rule 5 of the SM2 algorithm that is not being executed at all for > > cards > > graded 0 or 1 > > But rule 6 says not to apply rule 5 for failed cards: > > "If the quality response was lower than 3 then start repetitions for the item > from the beginning without changing the E-Factor"
It doesn't say that. It says that rule 6 applies only to cards with grades 0 and 1. It doesn't and shouldn't say anything about the applicability of rule 5 in rule 6! That information belongs in rule 5, which says nothing about treating differently graded cards differently, apart from the value of the EF increment. In fact, the page with the algorithm that you linked to has a link at the bottom to an example implementation. In that implementation, rule 5 is applied to all cards, regardless of the grade. Mnemosyne is really not using rule 5 correctly. > I guess the reasoning behind this was that after a lot a repetitions, > increasing the difficulty as well as resetting the interval was considered > too big of a penalty. I agree in the corner case you mention (immediately > failure), this seams suboptimal, but remember that this is not an exact > science, and the idea is that after many repetitions and corrections by the > user, the intervals and easiness factors converge to something which is > roughly OK. Failing to recall it after a lot of repetitions indicates that the intervals were too long. Starting again without decreasing the easiness will just repeat this. Since the intervals grow very rapidly anyway (exponentially), an EF that is a bit too low is not very problematic. However, a failure to recall a card is very serious and carries a very serious penalty: the interval is reset to 1 day. That's much worse than a slightly lower EF. So learning will be more efficient if a card is rated a bit too difficult, and never forgotten, rather than a bit too easy, and occasionally forgotten. Since, as you say, the EF should converge to a roughly OK value in the very long term anyway, the penalty to the EF that is incurred by a failure to recall has no serious consequences for the interval lengths in the very long term. But, after 0 or 1, the card has to be learned again from scratch. Once the intervals have grown again, the EF will converge. The EF penalty that goes with the reset therefore will only affect the short and intermediate term intervals. And that is where it improves the learning efficiency. > I'm hesitant to change the scheduler now after so many years without detailed > statistical analysis to back up any change. The data is there in the > collected learning logs, but analyzing it has not yet made it to the top of > my list. By fixing the scheduler now and implementing the algorithm as designed, you could actually get some interesting statistics for a comparison. Of course, putting a comment in the Mnemosyne documentation to explain the situation is much easier than analysing the data. With the current scheduler you really cannot claim that you are using SM2 as you are doing now in the docs, and you cannot make that claim either in any scientific publication based on the data that comes out of Mnemosyne. But if that is ok, because the precise algorithm is not important for what you want to do with this data, then there is no reason not to fix the scheduler now. Best regards, Astrid -- To UNSUBSCRIBE, email to [email protected] with a subject of "unsubscribe". Trouble? Contact [email protected]
